Prediction of properties from simulations: a re-examination with modern statistical methods

Area of Science:

• Computational chemistry
• Statistical modeling
• Drug discovery

Background:

• Existing linear regression models for predicting solvation free energies and partition equilibria have limitations.
• The Duffy and Jorgensen dataset presents challenges like discrepant observations and response curvature.
• Advanced statistical techniques are needed to improve predictive accuracy for molecular properties.

Purpose of the Study:

• To evaluate recently developed statistical models for predicting solvation free energies and partition equilibria.
• To compare the performance of generalized additive models (GAMs) and robust regression against traditional linear models.
• To identify improved methods for handling complex datasets with discrepant observations and curvature.

Main Methods:

• Application of robust parameter estimation to downweight influential observations.
• Model selection using linear representations, cubic polynomials, B-splines, and GAMs.
• Variable selection via formal tests and resampling methods like bootstrapping for prediction error assessment.

Main Results:

• Generalized additive models (GAMs) demonstrate superior performance compared to linear models for data description and prediction.
• Robust regression models and GAMs achieved the lowest conditional expected loss of prediction across four responses.
• Robust regression models effectively identified poorly fitting molecules, and prediction loss was reduced by approximately 50%.

Conclusions:

• Generalized additive models (GAMs) and robust regression offer significant improvements in predicting molecular properties.
• Robust regression aids in identifying problematic molecular data points, enhancing model reliability.
• Bootstrapping is a more reliable method than cross-validation for comparing predictive models in this context.